Flags and settings.
In [1]:
SAVE_FIGURES = False
PAPER_FEATURES = ['frequency', 'aoa', 'clustering', 'letters_count',
'synonyms_count', 'orthographic_density']
BIN_COUNT = 4
Imports and database setup.
In [2]:
import pandas as pd
import seaborn as sb
import numpy as np
%matplotlib inline
import matplotlib.pyplot as plt
from progressbar import ProgressBar
from statsmodels.stats.proportion import multinomial_proportions_confint
%cd -q ..
from brainscopypaste.conf import settings
%cd -q notebooks
from brainscopypaste.mine import Model, Time, Source, Past, Durl
from brainscopypaste.db import Substitution
from brainscopypaste.utils import init_db, session_scope, stopwords
engine = init_db()
Build our data.
In [3]:
def qposition(values, position):
value = values[position]
if np.isnan(value):
return np.nan, np.nan
finite_values = values[np.isfinite(values)]
svalues = np.array(sorted(finite_values))
length = len(svalues)
ours = np.where(svalues == value)[0]
return ours[0] / length, (ours[-1] + 1) / length
In [4]:
model = Model(time=Time.continuous, source=Source.all, past=Past.last_bin, durl=Durl.all, max_distance=1)
stop_poses = ['C', 'F', 'I', 'M', 'P', 'S', 'U']
data = []
# First get the exact substitution ids so we can get a working progress bar
# in the next step.
with session_scope() as session:
substitutions = session.query(Substitution.id)\
.filter(Substitution.model == model)
print("Got {} substitutions for model {}"
.format(substitutions.count(), model))
substitution_ids = [id for (id,) in substitutions]
for substitution_id in ProgressBar(term_width=80)(substitution_ids):
with session_scope() as session:
substitution = session.query(Substitution).get(substitution_id)
# Prepare these arrays for use in stopword-checking.
dslice = slice(substitution.start,
substitution.start
+ len(substitution.destination.tokens))
lemmas = substitution.source.lemmas[dslice]
tokens = substitution.source.tokens[dslice]
tags = substitution.source.tags[dslice]
is_stopword = np.array([(lemma in stopwords)
or (token in stopwords)
for (lemma, token) in zip(lemmas, tokens)])
for feature in Substitution.__features__:
# Get feature values for the sentence and its words.
sentence_values, _ = substitution.\
source_destination_features(feature)
sentence_values_rel, _ = substitution.\
source_destination_features(feature,
sentence_relative='median')
source_type, _ = Substitution.__features__[feature]
words = getattr(substitution.source, source_type)[dslice]
# Find the bins we'll use.
# If there are only NaNs or only one feature value
# we can't get bins on this sentence, so we want at least
# 2 different feature values.
# We also skip feature values if the source word is not coded
# for the feature, as it would skew the 'appeared'
# distributions compared to the distribution of substituted
# words. (For instance, the sum of categories would not be
# equal to the sum of H0s in the very last graphs,
# on sentencequantile. It also lets us make meaningful H0
# comparison in all the other feature-based graphs.)
non_sw_values = sentence_values.copy()
non_sw_values[is_stopword] = np.nan
non_sw_value_set = \
set(non_sw_values[np.isfinite(non_sw_values)])
if (len(non_sw_value_set) <= 1 or
np.isnan(sentence_values[substitution.position])):
allnans = [np.nan] * len(non_sw_values)
bins = allnans
non_sw_values = allnans
sentence_values = allnans
sentence_values_rel = allnans
else:
bins = pd.cut(non_sw_values, BIN_COUNT, labels=False)
# For each non-stopword, store its various properties.
for i, (word, tag, skip) in enumerate(zip(words, tags,
is_stopword)):
if skip:
# Drop any stopwords.
continue
# Get a readable POS tag
rtag = tag[0]
rtag = 'Stopword-like' if rtag in stop_poses else rtag
# Get the word's quantile position.
start_quantile, stop_quantile = qposition(non_sw_values, i)
# Store the word's properties.
data.append({
'cluster_id': substitution.source.cluster.sid,
'destination_id': substitution.destination.sid,
'occurrence': substitution.occurrence,
'source_id': substitution.source.sid,
'position': substitution.position,
'feature': feature,
'word': word,
'POS': tag,
'rPOS': rtag,
'target': i == substitution.position,
'value': sentence_values[i],
'value_rel': sentence_values_rel[i],
'bin': bins[i],
'start_quantile': start_quantile,
'stop_quantile': stop_quantile,
'word_position': i
})
words = pd.DataFrame(data)
del data
Assign proper weight to each substitution.
In [5]:
divide_target_all_sum = \
lambda x: x / (words.loc[x.index].target
* words.loc[x.index].weight_all).sum()
divide_target_feature_sum = \
lambda x: x / (words.loc[x.index].target
* words.loc[x.index].weight_feature).sum()
# Weight is 1, at first (or 1 for feature-coded substitutions).
words['weight_all'] = 1
words['weight_feature'] = 1 * np.isfinite(words.value)
# Divided by the number of substitutions that share a durl.
print('Computing shared durl (all) weights')
words['weight_all'] = words\
.groupby(['destination_id', 'occurrence', 'position',
'feature'])['weight_all']\
.transform(divide_target_all_sum)
print('Computing shared durl (per-feature) weights')
words['weight_feature'] = words\
.groupby(['destination_id', 'occurrence', 'position',
'feature'])['weight_feature']\
.transform(divide_target_feature_sum)
# Divided by the number of substitutions that share a cluster.
# (Using divide_target_sum, where we divide by the sum of weights,
# ensures we count only one for each group of substitutions sharing
# a same durl.)
print('Computing shared cluster (all) weights')
words['weight_all'] = words\
.groupby(['cluster_id', 'feature'])['weight_all']\
.transform(divide_target_all_sum)
print('Computing shared cluster (per-feature) weights')
words['weight_feature'] = words\
.groupby(['cluster_id', 'feature'])['weight_feature']\
.transform(divide_target_feature_sum)
# Add a weight measure for word appearances, weighing a word
# by the number of words that appear with it in its sentence.
# And the same for substitutions *whose source is coded by the feature*.
# (This lets us have the sum of categories equal the sum of H0s
# in the very last graphs [on sentencequantile], and make meaningful H0
# comparison values for all the other feature-based graphs.)
print('Computing appeared (all) weights')
words['weight_all_appeared'] = words\
.groupby(['source_id', 'destination_id', 'occurrence',
'position', 'feature'])['weight_all']\
.transform(lambda x: x / len(x))
print('Computing appeared (per-feature) weights')
words['weight_feature_appeared'] = words\
.groupby(['source_id', 'destination_id', 'occurrence',
'position', 'feature'])['weight_feature']\
.transform(lambda x: x / np.isfinite(words.loc[x.index].value).sum())
# In the above, note that when using a model that allows for multiple
# substitutions, those are stored as two separate substitutions in the
# database. This is ok, since we count the number of times a word is
# substituted compared to what it would have been substituted at
# random (i.e. we measure a bias, not a probability). Which leads us to
# count multiple substitutions in a same sentence as *different*
# substitutions, and to reflect this in the weights we must group
# substitutions by the position of the substituted word also (which is
# what we do here).
Prepare feature ordering.
In [6]:
ordered_features = sorted(
Substitution.__features__,
key=lambda f: Substitution._transformed_feature(f).__doc__
)
Prepare counting functions.
In [7]:
target_all_counts = \
lambda x: (x * words.loc[x.index, 'weight_all']).sum()
target_feature_counts = \
lambda x: (x * words.loc[x.index, 'weight_feature']).sum()
appeared_all_counts = \
lambda x: words.loc[x.index, 'weight_all_appeared'].sum()
appeared_feature_counts = \
lambda x: words.loc[x.index, 'weight_feature_appeared'].sum()
susty_all = \
lambda x: target_all_counts(x) / appeared_all_counts(x)
susty_feature = \
lambda x: target_feature_counts(x) / appeared_feature_counts(x)
In [8]:
# Compute POS counts.
susties_pos = words[words.feature == 'aoa']\
.groupby('rPOS')['target']\
.aggregate({'susceptibility': susty_all,
'n_substituted': target_all_counts,
'n_appeared': appeared_all_counts})\
.rename_axis('POS group')
# Plot.
fig, axes = plt.subplots(2, 1, figsize=(8, 8))
# Raw substituted and appeared values.
susties_pos[['n_substituted', 'n_appeared']]\
.plot(ax=axes[0], kind='bar', rot=0)
# With their CIs.
total_substituted = susties_pos.n_substituted.sum()
cis = multinomial_proportions_confint(susties_pos.n_substituted.round(),
method='goodman')
for i in range(len(susties_pos)):
axes[0].plot([i-.125, i-.125], cis[i] * total_substituted,
lw=4, color='grey',
label='95% CI' if i == 0 else None)
axes[0].legend()
# Substitutability values.
susties_pos['susceptibility']\
.plot(ax=axes[1], kind='bar', legend=True, ylim=(0, 2), rot=0)
axes[1].set_ylabel(r'$susceptibility = \frac{substituted}{appeared}$')
# With their CIs.
for i in range(len(susties_pos)):
axes[1].plot([i, i], (cis[i] * total_substituted
/ susties_pos.n_appeared.iloc[i]),
lw=4, color='grey',
label='95% CI' if i == 0 else None)
axes[1].legend(loc='best')
# Save if necessary.
if SAVE_FIGURES:
fig.savefig(settings.FIGURE.format('all-susceptibilities-pos'),
bbox_inches='tight', dpi=300)
Note on confidence intervals
Here we're in case (3) of the explanation below on confidence intervals (in section 3): it's really like a multinomial sampling, but not quite since not all POS tags are available to sample from in all the sentences. There's no way out of this, so we're going to use multinomial CIs. We can safely scale all the bars and CIs to their respective n_appeared values, since that is an independent given before the sampling.
Are the appeared and substituted proportions statistically different?
The only test we can easily do is a multinomial goodness-of-fit. This tells us if the n_substituted counts are significantly different from the reference n_appeared counts.
From there on we know a few things:
n_substituted count to its reference n_appeared count tells us if it's statistically different (< or >). We know this will be true individually for any POS that is out of its confidence region for the global goodness-of-fit test, since it's a weaker hypothesis (so the null rejection region will be wider, and the POS we're looking at is already in the rejection region for the global test). We don't know if it'll be true or not for POSes that are in their confidence region for the global test.n_substituted counts to their reference n_appeared counts tells us if there is bias for one w.r.t. the other. This is also true for all pairs of POSes that are on alternate sides of their confidence region in the global test (for the same reasons as in the previous point). We don't know if it's true for the other POSes though.
In [9]:
# Test the n_substituted proportions are different from
# the n_appeared proportions
total_appeared = susties_pos.n_appeared.sum()
appeared_cis = multinomial_proportions_confint(
susties_pos.n_appeared.round(), method='goodman')
differences = [(s < ci[0] * total_appeared) or (s > ci[1] * total_appeared)
for s, ci in zip(susties_pos.n_substituted, appeared_cis)]
are_different = np.any(differences)
if are_different:
print("Appeared and substituted proportions are different with p < .05")
print("The following POS tags are out of their confidence region:",
list(susties_pos.index[np.where(differences)[0]]))
else:
print("Appeared and substituted proportions cannot be "
"said different with p value better than .05")
Prepare plotting functions, for bin and quartile susceptibilities for each feature.
In [10]:
def print_significance(feature, h0s, heights):
h0_total = h0s.sum()
bin_count = len(h0s)
print()
print('-' * len(feature))
print(feature)
print('-' * len(feature))
for n_stars, alpha in enumerate([.05, .01, .001]):
h0_cis = multinomial_proportions_confint(h0s.round(),
method='goodman',
alpha=alpha)
differences = ((heights < h0_cis[:, 0] * h0_total)
| (heights > h0_cis[:, 1] * h0_total))
are_different = np.any(differences)
stars = ' ' * (3 - n_stars) + '*' * (1 + n_stars)
if are_different:
bins_different = np.where(differences)[0]
bins_different += np.ones_like(bins_different)
print(stars + ' Target different H_0 with p < {}.'
' Bins [1; {}] out of region: {}'
.format(alpha, bin_count, bins_different.tolist()))
else:
print(' Target NOT different from H_0 (p > {})'
.format(alpha))
break
In [11]:
def plot_bin_susties(**kwargs):
data = kwargs['data']
feature = data.iloc[0].feature
color = kwargs.get('color', 'blue')
relative = kwargs.get('relative', False)
quantiles = kwargs.get('quantiles', False)
value = data.value_rel if relative else data.value
# Compute binning.
cut, cut_kws = ((pd.qcut, {}) if quantiles
else (pd.cut, {'right': False}))
for bin_count in range(BIN_COUNT, 0, -1):
try:
value_bins, bins = cut(value, bin_count, labels=False,
retbins=True, **cut_kws)
break
except ValueError:
pass
middles = (bins[:-1] + bins[1:]) / 2
# Compute bin counts. Note here the bins are computed on the
# distribution of observed substitutions, not the simulated aggregated
# distributions of cluster-unit substitutions. But since it's mostly
# deduplication that the aggregation process addresses, the bins
# should be mostly the same. This could be corrected by computing
# bins on the aggregate distribution (not hard), but it's really
# not important now.
heights = np.zeros(bin_count)
h0s = np.zeros(bin_count)
for i in range(bin_count):
heights[i] = (data[data.target & (value_bins == i)]
.weight_feature.sum())
h0s[i] = data[value_bins == i].weight_feature_appeared.sum()
total = sum(heights)
cis = (multinomial_proportions_confint(heights.round(),
method='goodman')
* total / h0s[:, np.newaxis])
# Plot them.
sigmaphi = (r'\sigma_{\phi'
+ ('_r' if relative else '')
+ '}')
plt.plot(middles, heights / h0s,
color=color, label='${}$'.format(sigmaphi))
plt.fill_between(middles, cis[:, 0], cis[:, 1],
color=sb.desaturate(color, 0.2), alpha=0.2)
plt.plot(middles, np.ones_like(middles), '--',
color=sb.desaturate(color, 0.2),
label='${}^0$'.format(sigmaphi))
plt.xlim(middles[0], middles[-1])
plt.ylim(0, 2)
# Test for statistical significance
print_significance(feature, h0s, heights)
In [12]:
def plot_grid(data, features, filename,
plot_function, xlabel, ylabel, plot_kws={}):
g = sb.FacetGrid(data=data[data['feature']
.map(lambda f: f in features)],
sharex=False, sharey=True,
col='feature', hue='feature',
col_order=features, hue_order=features,
col_wrap=3, aspect=1.5, size=3)
g.map_dataframe(plot_function, **plot_kws)
g.set_titles('{col_name}')
g.set_xlabels(xlabel)
g.set_ylabels(ylabel)
for ax in g.axes.ravel():
legend = ax.legend(frameon=True, loc='best')
if not legend:
# Skip if nothing was plotted on these axes.
continue
frame = legend.get_frame()
frame.set_facecolor('#f2f2f2')
frame.set_edgecolor('#000000')
ax.set_title(Substitution._transformed_feature(ax.get_title())
.__doc__)
if SAVE_FIGURES:
g.fig.savefig(settings.FIGURE.format(filename),
bbox_inches='tight', dpi=300)
In [13]:
plot_grid(words, ordered_features,
'all-susceptibilities-fixedbins_global',
plot_bin_susties, r'$\phi$', 'Susceptibility')
Note on how graphs and their confidence intervals are computed here
There are three ways I can do a computation like above:
(1) For each word, we look at how many times it is substituted versus how many times it appears in a position where it could have been substituted. This is the word's susceptibility, $\sigma(w)$. Then for each feature bin $b_i$ we take all the words such that $\phi(w) \in b_i$, average, and compute an asymptotic confidence interval based on how many words are in the bin. This fails for sentence-relative features, because a given word has different feature values depending on the sentence it appears in. So we discard this.
(2) Do the same but at the feature value level. So we define a feature value susceptibility, $\sigma_{\phi}(f)$, and compute a confidence interval based on how many different feature values we have in the bin. The idea behind (1) and (2) is to look at the bin middle-value like the relevant object we're measuring, and we have several measures for each bin middle-value, hence the confidence interval. In each bin $b_i$ we have:
$$\left< \sigma_{\phi}(f) \right>_{f \in b_i}$$The problem with both (1) and (2) is that there's no proper $\mathcal{H}_0$ value, because the averages in the bins don't necessarily equal 1 under $\mathcal{H}_0$. Also, we can't check that there is consistency, showing that the sum of susceptibility values of the bins is 1. Hence case 3:
(3) Consider that we sample a multinomial process: each substitution is in fact the sampling of a feature value from one of the four bins. In that case, we can compute multinomial proportion CIs. This is also not completely satisfactory since in most cases not all feature values are available at the time of sampling, since most sentences don't range over all the feature's values, but it's what lets us compute proper null hypotheses: in each bin $b_i$ we have a value of $\sigma_{\phi}(b_i)$, and the sum of those should be the same under $\mathcal{H}_0$ as in the experiment (in practice in the graphs, we divide by the values under $\mathcal{H}_0$, and the reference is $\sigma_{\phi}^0(b_i) = 1$).
Here and below, we're always in case (3).
In [14]:
plot_grid(words[~(((words.feature == 'letters_count')
& (words.value > 15))
| ((words.feature == 'aoa')
& (words.value > 15))
| ((words.feature == 'clustering')
& (words.value > -3)))],
PAPER_FEATURES,
'paper-susceptibilities-fixedbins_global',
plot_bin_susties, r'$\phi$', 'Susceptibility')
In [15]:
plot_grid(words, ordered_features,
'all-susceptibilities-quantilebins_global', plot_bin_susties,
r'$\phi$', 'Susceptibility',
plot_kws={'quantiles': True})
Note on confidence intervals
Here we're again in case (2) of the above explanation on confidence intervals (in section 3.1), since we're just binning by quantiles instead of fixed-width bins.
In [16]:
plot_grid(words, PAPER_FEATURES,
'paper-susceptibilities-quantilebins_global', plot_bin_susties,
r'$\phi$', 'Susceptibility',
plot_kws={'quantiles': True})
In [17]:
plot_grid(words, ordered_features,
'all-susceptibilities-fixedbins_sentencerel',
plot_bin_susties, r'$\phi_r$', 'Susceptibility',
plot_kws={'relative': True})
In [18]:
plot_grid(words, PAPER_FEATURES,
'paper-susceptibilities-fixedbins_sentencerel',
plot_bin_susties, r'$\phi_r$', 'Susceptibility',
plot_kws={'relative': True})
In [19]:
plot_grid(words, ordered_features,
'all-susceptibilities-quantilebins_sentencerel',
plot_bin_susties, r'$\phi_r$', 'Susceptibility',
plot_kws={'quantiles': True, 'relative': True})
In [20]:
plot_grid(words, PAPER_FEATURES,
'paper-susceptibilities-quantilebins_sentencerel',
plot_bin_susties, r'$\phi_r$', 'Susceptibility',
plot_kws={'quantiles': True, 'relative': True})
In [21]:
def plot_sentencebin_susties(**kwargs):
data = kwargs['data']
color = kwargs.get('color', 'blue')
feature = data.iloc[0].feature
# Compute bin counts
heights = np.zeros(BIN_COUNT)
h0s = np.zeros(BIN_COUNT)
for i in range(BIN_COUNT):
heights[i] = (data[data.target & (data.bin == i)]
.weight_feature.sum())
h0s[i] = data[data.bin == i].weight_feature_appeared.sum()
total = sum(heights)
cis = (multinomial_proportions_confint(heights.round(),
method='goodman')
* total / h0s[:, np.newaxis])
# Plot them.
sigmaphi = r'\sigma_{bin_{\phi}}'
x = range(1, BIN_COUNT + 1)
plt.plot(x, heights / h0s, color=color, label='${}$'.format(sigmaphi))
plt.fill_between(x, cis[:, 0], cis[:, 1],
color=sb.desaturate(color, 0.2), alpha=0.2)
plt.plot(x, np.ones_like(x), '--',
color=sb.desaturate(color, 0.2),
label='${}^0$'.format(sigmaphi))
plt.xticks(x)
plt.ylim(0, None)
# Test for significance.
print_significance(feature, h0s, heights)
In [22]:
plot_grid(words, ordered_features,
'all-susceptibilities-sentencebins',
plot_sentencebin_susties, r'$bin_{\phi}$ in sentence',
'Susceptibility')
In [23]:
plot_grid(words, PAPER_FEATURES,
'paper-susceptibilities-sentencebins',
plot_sentencebin_susties, r'$bin_{\phi}$ in sentence',
'Susceptibility')
For each feature, count the sum of weights in each bin and plot that.
In [24]:
def bound(limits, values):
left, right = limits
assert left < right
return np.maximum(left, np.minimum(right, values))
In [25]:
def plot_sentencequantile_susties(**kwargs):
data = kwargs['data']
color = kwargs.get('color', 'blue')
feature = data.iloc[0].feature
# Compute bin counts
heights = np.zeros(BIN_COUNT)
h0s = np.zeros(BIN_COUNT)
step = 1 / BIN_COUNT
for i in range(BIN_COUNT):
limits = [i * step, (i + 1) * step]
contributions = ((bound(limits, data.stop_quantile)
- bound(limits, data.start_quantile))
/ (data.stop_quantile - data.start_quantile))
heights[i] = \
(contributions * data.weight_feature)[data.target].sum()
h0s[i] = (contributions * data.weight_feature_appeared).sum()
total = sum(heights)
cis = (multinomial_proportions_confint(heights.round(),
method='goodman')
* total)# / h0s[:, np.newaxis])
# Plot them.
sigmaphi = r'\sigma_{q_{\phi}}'
x = range(1, BIN_COUNT + 1)
plt.plot(x, heights,# / h0s,
color=color, label='${}$'.format(sigmaphi))
plt.fill_between(x, cis[:, 0], cis[:, 1],
color=sb.desaturate(color, 0.2), alpha=0.2)
plt.plot(x, h0s, '--',
color=sb.desaturate(color, 0.2),
label='${}^0$'.format(sigmaphi))
plt.xticks(x)
plt.ylim(0, None)
# Test for significance.
print_significance(feature, h0s, heights)
In [26]:
plot_grid(words, ordered_features,
'all-susceptibilities-sentencequantiles',
plot_sentencequantile_susties, r'$q_{\phi}$ in sentence',
'Number of substitutions\n(weighted to cluster unit)')
In [27]:
plot_grid(words, PAPER_FEATURES,
'paper-susceptibilities-sentencequantiles',
plot_sentencequantile_susties, r'$q_{\phi}$ in sentence',
'Number of substitutions\n(weighted to cluster unit)')
We try to predict which words are substituted, based on their global values, sentence-relative values, bins and quantiles of those, or in-sentence bin values.
Prediction is not good, mainly because the constraint of one-substitution-per-sentence can't be factored in the model simply. So precision is generally very low, around .20-.25, and when accuracy goes up recall plummets.
So it might show some interaction effects, but given that the fit is very bad I wouldn't trust it.
In-sentence quantiles (from section 5.2) were not done, as they're impossible to reduce to one value (our measure of those quantiles is in fact a subrange of [0, 1] for each word, corresponding to the subrange of the sentence distribution that that word's feature value represented).
In [28]:
from sklearn.linear_model import LogisticRegression
from sklearn.preprocessing import PolynomialFeatures
from scipy.stats import binom_test
In [29]:
def regress_binning(data, features, value_funcs):
# Compute bins
data = data.copy()
regress_features = [('{}'.format(value_name), feature)
for value_name in value_funcs.keys()
for feature in features]
for i, (value_name, value_func) in enumerate(value_funcs.items()):
data[value_name] = value_func(data)
# Massage the dataframe to have feature bin as columns.
data_wide = pd.pivot_table(
data,
values=list(value_funcs.keys()),
index=['destination_id', 'occurrence', 'source_id', 'position',
'word_position'],
columns=['feature']
)[regress_features]
# Add the target value.
# Question/FIXME: should we use weight_appeared for regression?
data_wide['target'] = pd.pivot_table(
data,
values=['target'],
index=['destination_id', 'occurrence', 'source_id', 'position',
'word_position'],
columns=['feature']
)[('target', 'aoa')]
data_wide = data_wide.dropna()
# Compute polynomial features.
poly = PolynomialFeatures(degree=2, interaction_only=True)
pdata = poly.fit_transform(data_wide[regress_features])
pregress_features = [' * '.join(['_'.join(regress_features[j])
for j, p in enumerate(powers)
if p > 0]) or 'intercept'
for powers in poly.powers_]
# Divide into two sets.
print('Regressing with {} word measures (divided into'
' training and prediction sets)'
.format(len(data_wide)))
pdata_train = pdata[:len(data_wide) // 2]
target_train = data_wide.iloc[:len(data_wide) // 2].target
pdata_predict = pdata[len(data_wide) // 2:]
target_predict = data_wide.iloc[len(data_wide) // 2:].target
assert len(pdata_train) + len(pdata_predict) == len(data_wide)
assert len(target_train) + len(target_predict) == len(data_wide)
# Regress
regressor = LogisticRegression(penalty='l2', class_weight='balanced',
fit_intercept=False)
regressor.fit(pdata_train, target_train)
# And predict
prediction = regressor.predict(pdata_predict)
standard = target_predict.values
success = prediction == standard
tp = prediction & standard
tn = (~prediction) & (~standard)
fp = prediction & (~standard)
fn = (~prediction) & standard
print()
print('{:.2f}% of words well predicted (non-random at p = {:.1})'
.format(100 * success.mean(),
binom_test(success.sum(), len(success))))
print('Precision = {:.2f}'.format(standard[prediction].mean()))
print('Recall = {:.2f}'.format(prediction[standard].mean()))
print()
print('Coefficients:')
print(pd.Series(index=pregress_features, data=regressor.coef_[0]))
Global feature value
In [30]:
regress_binning(words, ['frequency', 'aoa', 'letters_count',
'orthographic_density'],
{'global': lambda d: d.value})
Sentence-relative feature value
In [31]:
regress_binning(words, ['frequency', 'aoa', 'letters_count',
'orthographic_density'],
{'sentence-rel': lambda d: d.value_rel})
Global + sentence-relative feature values
In [32]:
regress_binning(words, ['frequency', 'aoa', 'letters_count',
'orthographic_density'],
{'global': lambda d: d.value,
'sentence-rel': lambda d: d.value_rel})
(3.1) Bins of distribution of appeared global feature values
In [33]:
regress_binning(words, ['frequency', 'aoa', 'letters_count',
'orthographic_density'],
{'bins-global':
lambda d: pd.cut(d.value, BIN_COUNT,
labels=False, right=False)})
(3.2) Quantiles of distribution of appeared global feature values
In [34]:
regress_binning(words, ['frequency', 'aoa', 'letters_count',
'orthographic_density'],
{'quantiles-global':
lambda d: pd.qcut(d.value, BIN_COUNT, labels=False)})
(4.1) Bins of distribution of appeared sentence-relative values
In [35]:
regress_binning(words, ['frequency', 'aoa', 'letters_count',
'orthographic_density'],
{'bins-sentence-rel':
lambda d: pd.cut(d.value_rel, BIN_COUNT,
labels=False, right=False)})
(4.2) Quantiles of distribution of appeared sentence-relative values
In [36]:
regress_binning(words, ['frequency', 'aoa', 'letters_count',
'orthographic_density'],
{'quantiles-sentence-rel':
lambda d: pd.qcut(d.value_rel, BIN_COUNT,
labels=False)})
(5.1) In-sentence bins (of distribution of values in each sentence)
In [37]:
regress_binning(words, ['frequency', 'aoa', 'letters_count',
'orthographic_density'],
{'in-sentence-bins': lambda d: d.bin})
In [38]:
from sklearn.decomposition import PCA
In [39]:
def pca_values(data, features, value_func):
data = data.copy()
data['pca_value'] = value_func(data)
# Prepare dataframe, averaging over shared durl.
data_wide = pd.pivot_table(
data[data.target],
values='pca_value',
index=['cluster_id', 'destination_id', 'occurrence',
'position'],
columns=['feature']
)[features]
# ... then over shared clusters, and dropping NaNs.
data_wide = data_wide\
.groupby(level='cluster_id')\
.agg(np.mean)\
.dropna(how='any')
print('Computing PCA on {} aggregated word measures'
.format(len(data_wide)))
print()
# Compute PCA.
pca = PCA(n_components='mle')
pca.fit(data_wide)
print('Variance explained by first {} components (mle-estimated): {}'
.format(pca.n_components_, pca.explained_variance_ratio_))
print()
print('Components:')
print(pd.DataFrame(index=data_wide.columns,
data=pca.components_.T,
columns=['Comp. {}'.format(i)
for i in range(pca.n_components_)]))
PCA of feature value of substituted words
In [40]:
pca_values(words, ['frequency', 'aoa', 'letters_count'],
lambda d: d.value)
PCA of sentence-relative value of substituted words
In [41]:
pca_values(words, ['frequency', 'aoa', 'letters_count'],
lambda d: d.value_rel)